To achieve proactive diabetes management in type 1 diabetes, accurate glucose prediction is critical. Due to complex glucose dynamics, diabetic glucose levels often exhibit highly nonlinear and nonstationary patterns, making accurate prediction challenging. Existing methods mainly focus on analyzing glucose levels in the time domain, failing to capture complex patterns. To address these gaps, a multi-scale time-frequency approach is proposed for accurate and personalized glucose prediction called MTGlu. A time-frequency decomposition-based module with optimized segmentation is proposed to uncover multi-scale nonstationary oscillation patterns. The temporal dependency learning module is further developed to capture nonlinear patterns at each scale. Moreover, we propose a new evaluation metric, abnormal time percentage, which accounts for the impact of abnormal glucose predictions on treatment. The proposed approach performs well on real patient data, achieving a 42% improvement in prediction performance on average compared to other methods.